With President Joe Biden’s executive order on artificial intelligence (AI) this week emphasizing U.S. leadership in the field and Apple touting the utility to AI developers of the most powerful new laptop computers it launched, Beijing is not standing idly by.
As part of China’s effort to advance technological innovation and self-reliance, the government is undertaking a massive plan to build out the domestic computing ecosystem encompassing data processing, storage and transmission.
The Action Plan for the High-Quality Development of Computing Infrastructure will significantly enhance computing infrastructure to serve the development of China’s cloud and AI sectors.
It was issued by six authorities, including the Ministry of Industry and Information Technology (MIIT), the Ministry of Education and the People’s Bank of China, underlining its national importance.
Specifically, the plan intends to realize an infrastructure with supercomputers capable of processing over 300 exaFLOPS by 2025.
In August, the MIIT said that China’s total supercomputing power had reached 197 exaFLOPS, ranking second globally behind the United States, with around 230 exaFLOPs.
Thus, the planned expansion is a geopolitical and technological goal.
An exaFLOP is 1 quintillion floating point operations per second. Be assured that that is a lot of computing performance: the world’s fastest supercomputer, HP Cray’s Frontier, clocks in at 1.2 exaFLOPS. China’s target implies more and more powerful supercomputers than it now has.
In June, the TOPS500 project at the University of Tennessee at Knoxville, which keeps score, counted 130 to 150 in the United States. China is believed to have prototype exaFLOP systems.
One-third of its targeted 2025 computing power will be graphics processors (GPUs) rather than traditional computer chips (CPUs). GPUs have far more cores than CPUs and are thus able to perform far more tasks in parallel.
Beijing will be less interested in their capacity to render the increasingly realistic graphics of computer games than in the fact that parallel processes accelerate training and inference for the highly computing-intensive models underpinning deep learning in AI and which enable general-purpose AI (GenAI) applications such as OpenAI’s hugely successful ChatGPT.
The plan further specifies a national data storage capability of 1,800 exabytes, up from 725 exabytes in 2022, of which 30% should be “advanced storage,” such as flash drives rather than traditional hard drives.
Flash drives’ faster speed makes them more suitable for AI applications. Blu-ray technology, best remembered in the West as a failed successor to the DVD, is envisaged as an option for the “cold storage” of data not often used.
The plan also envisages the completion of a national data grid to high latency standards for fast data transfers and the establishment of specialized infrastructure for integrating AI into sectors ranging from education to finance, transport, healthcare, energy and manufacturing.
One primary objective is to cut the cost of computing power for AI start-ups.
Running the supporting infrastructure for AI is expensive. For instance, ChatGPT currently costs USD700,000 per day to run. If it were to grow to receive one-tenth of the queries that Google’s search engine processes, it would require an immediate investment of over USD48bn and a recurring spend of USD16bn on chips alone.
These high costs are good news for Chinese chip manufacturers and incentivize them to develop their GPU capabilities, but it would be prohibitive for promising infant businesses.
In response, the Beijing municipal government has proposed subsidies of up to 20% of computing costs for Beijing-based large-language-model businesses, up to a maximum of USD273,000.
This amount is neither meaningful to giants such as Baidu and Alibaba, who would burn through it in days, nor sufficient to develop highly sophisticated “frontier AI.”
However, it will be significant for smaller AI firms with targeted applications in the priority sectors. It will also enhance the adoption of AI more broadly in government and society, creating a fertile ecosystem and likely a few future winners—mainly through integrating AI applications and cloud services.
A significant challenge to the plan is environmental.
AI centers are energy intensive. A recent estimate suggested that if every Google search interaction used GenAI, the company’s energy consumption could increase to 29.3 terawatt hours (TWh) per year—the equivalent of Ireland’s annual consumption—from 18.3 TWh in 2021.
Beijing’s plan encourages the development of more efficient computing equipment, specifically more power-efficient semiconductors. However, it stops short of setting quantitative targets, probably due to uncertainty about access to future (Western) technology.
A large effort to move data centers from the densely populated Eastern seaboard to the Western hinterland is underway in pursuit of less expensive land and more plentiful solar and wind energy.
Geopolitically, China’s computing schemes will likely face headwinds, too.
A year after its original sweeping semiconductor sanctions, Washington further tightened its rules, meaning that Nvidia—the leading manufacturer of GPUs—could no longer lawfully sell its A800 and H800 GPUs in China. Competing chips from Intel’s Habana Labs, such as the Gaudi2, are also affected.
These chips were specially developed after the earlier sanctions to stay just below the sanction thresholds. As a result, although China could not acquire the most advanced GPUs, the ‘detuned’ alternatives remained powerful enough, at least for the current generation of GenAI models.
The AI Silk Road
The plan encourages Chinese AI firms to internationalize, particularly in countries associated with the Belt and Road Initiative.
Yet, those businesses will increasingly face difficulties in convincing their clients that they will continue to deliver in cases where they are reliant on U.S. components or that going with Chinese technology will not attract negative attention from Washington.